[Coin-discuss] Performance of CoinMemcpyN()
Lou Hafer
lou at cs.sfu.ca
Fri Jul 29 17:15:26 EDT 2005
I think the issue here is more the quality of compiler optimization.
To take CoinMemcpyN as the example, what you're seeing is an explicit 8-fold
loop unroll to explicitly expose potential parallelism, in case the compiler
might not find it on its own. One would hope that memcpy on a given
architecture would be finely tuned for that architecture, but that's not
always the case.
Perhaps one of the early developers remembers which architecture
and/or compiler prompted this bit of code?
Lou
More information about the Coin-discuss
mailing list